The problem
Generative AI arrived in education faster than anyone planned for, and it brings both opportunity and risk. Whether it helps a student or hollows out their work depends almost entirely on how it is used. What this survey shows first is where it entered their lives: not through entertainment, but through schoolwork.
Homework and assignments were the most commonly reported use-case, selected by 260 respondents, followed by exam preparation at 236. Writing and images/design were also common, selected by 157 and 156 respondents respectively. Entertainment, by contrast, was selected by only 73, and just 11 respondents said they do not use AI at all. Within this sample AI functions primarily as an academic assistant, not a toy.
That distinction matters. AI is being used inside assessment and learning — the two places where getting an answer wrong has consequences — which makes trust and dependence the questions worth asking.
The important policy question is not whether students have tried AI — it is whether they are beginning to rely on tools whose accuracy and limitations they may not fully understand.
The rest of this report tests that question directly. It examines whether more frequent use is associated with wider tool adoption, greater perceived reliability, and greater dependence on AI in academic work — and what students say is stopping them from using it better.
What the survey shows
AI use is concentrated in a small group of widely available tools. ChatGPT was the most common, named by 312 of the 355 respondents who answered this question, followed by Gemini with 269 and Claude with 227. Meta AI, Perplexity, and NotebookLM formed a second tier — 109, 86, and 78 respondents respectively. Because respondents could select several platforms, the 355 of them account for 1,268 selections: an average of 3.54 tools each. Very few students rely on a single one.
As shown in Fig. 3, students overwhelmingly learn to use AI on their own: 336 of 358 selected this option, against 63 who credited social media, 37 YouTube, and 13 a paid course. Just 22 — one student in sixteen — learned it in school or an educational workspace. Students are adopting AI far faster than institutions are teaching it, which is what makes accuracy, responsible use, and academic dependence pressing rather than theoretical.
Does frequent use increase trust and dependence?
For many young people, AI is no longer just a tool to experiment with — it has become part of everyday studying, problem-solving and creating. But as AI becomes more familiar, does greater confidence in its reliability also translate into greater dependence? To explore this, we examined respondents’ perceived AI reliability alongside their self-reported level of dependence on AI.
Every one of the 358 respondents answered both scales, so this figure covers the whole sample. It shows a moderate positive association (ρ = 0.38, p < 0.001) between perceived reliability and dependence. Students who consider AI more reliable do tend to rely on it more, and with a sample this size that relationship is not a fluke.
The spread, however, is wide, and it is the more interesting part of the figure. Eighteen respondents sit in the top corner — 10 out of 10 on both scales — and one rates reliability 3 while reporting dependence of 10, leaning completely on a tool they say they do not trust. Others rate AI 9 or 10 for reliability and still report low dependence. Trust accounts for about 17% of the variation in dependence; the remaining five sixths comes from somewhere else.
Trust therefore encourages reliance without determining it. Convenience, habit, deadline pressure, and the kind of task at hand are all doing work this figure cannot separate out. That matters for policy: an intervention aimed only at correcting how much students trust AI would leave most of the reason they depend on it untouched.
Barriers to AI
What stops students from getting better at AI is not mainly the price of courses. It is time, and not knowing where to begin. As shown in Fig. 5, 101 respondents — 28.2% of the sample — selected “Nothing stops me,” so a substantial minority sees no obstacle at all. Among the rest, the pattern is consistent.
The most frequently selected difficulty was lack of time, reported by 95 respondents. Students may recognise the value of learning about AI but struggle to dedicate time to it alongside schoolwork, examinations, and other commitments. The next most common response was not knowing where to start, selected by 74 respondents — a notable finding given that students already use AI extensively, yet many appear to lack a clear pathway toward using it more effectively, critically, or responsibly.
Financial barriers appeared less significant than these practical concerns. “It costs money” was selected by 55 respondents, compared with only 33 who cited the absence of a good free course. The core problem is not a shortage of free resources, but the difficulty of identifying resources that are structured, credible, and relevant to students’ needs. It is worth noting alongside this that 69.3% of respondents have never paid for an AI tool: cost shapes which tools students reach for, even where it is not what stops them improving.
How does AI use deepen?
Frequency, breadth, trust and dependence do not move independently. Read together, the survey describes a progression — each stage making the next more likely.
Frequency widens the toolkit. The 177 students who use AI several times a day averaged 4.19 tools each, against 3.54 across the sample and 2.91 among everyone else. Using AI more does not mean using one tool harder; it means reaching for more of them.
Breadth raises trust. Those same heavy users rated AI 6.99 out of 10 for reliability, against 5.97 for the rest of the sample — a full point higher. Exposure appears to build confidence in the answers.
Trust turns into dependence. This is the association in Fig. 4 (ρ = 0.38, p < 0.001). Heavy users report dependence of 5.92 out of 10, against 4.09 for everyone else.
What blocks them changes. Newer users are stopped by not knowing where to begin — 74 respondents, the third most common answer. Established users are no longer looking for a starting point but for depth: harder tasks and credible guidance.
And it reaches their plans. 60.1% report that AI has changed the career they intend to pursue (31.6%) or left them unsure of it (28.5%). Only 39.9% say it has changed nothing.
The progression is the finding. A student does not decide once to depend on AI; they arrive at it by using it often enough that it becomes the obvious first move. Every stage above happens without a teacher present.
Policy implications
Trust in AI is linked to dependence, but it is not the only factor. Since convenience, habit, and the type of task also seem to influence reliance, policy should focus not only on improving trust but also on encouraging responsible use.
Educational institutions
- Introduce clear AI-use guidelines. Schools and colleges should explain when AI can be used and when it should not be. Students could be allowed to use AI for brainstorming or improving drafts, while independent work is required for assessments — separating AI as support from AI as dependence.
- Include AI-use awareness in digital literacy. Teach students to recognise when they are relying on AI too much. Simple exercises — completing a task independently before using AI, then comparing the two — build confidence in their own abilities while still capturing AI’s benefits.
AI developers & platforms
- Make AI limitations more visible. Platforms should clearly indicate when an answer may be uncertain and provide sources where possible, helping users develop a realistic level of trust rather than assuming AI is always correct.
- Encourage users to think before relying on AI. Simple reminders during repeated or heavy use — such as asking whether the user wants to try solving a problem themselves first — could reduce automatic dependence without preventing genuinely helpful use.
A note on the numbers
What this survey can and cannot support, stated plainly rather than left for the reader to work out.
Conclusion
Central, not supplementary
AI has already changed how the students in this sample study, check information, and think about what they will do for a living. What started as trying out a chatbot has become a daily, multi-tool routine that ordinary schoolwork now runs on. The more familiar it gets, the more they lean on it — until automated help is no longer something extra, but part of how the work gets done.
The consequences reach past the classroom. 336 of 358 respondents taught themselves, and 22 learned at school. That leaves a generation forming habits around accuracy, sourcing and judgement with nobody checking the habits — and 60.1% already redrawing their career plans around a technology they were never formally taught.
The instinct to respond with bans is understandable and, on this evidence, beside the point. Students are not waiting for permission; they are past that. The useful question is no longer whether schools should teach AI, but whether they can afford not to — because the alternative, which is what this survey documents, is students learning it alone.